Utter – Native macOS on-device voice dictation app powered by whisper.cpp and Metal

[h3x-levi]

Hi everyone! 👋

I have been working on Utter, a native on-device voice dictation app for macOS built with Swift and SwiftUI, powered by whisper.cpp.

I built this project completely solo to solve a personal pain point: cloud-based dictation tools upload audio to remote servers, require paid subscriptions, and introduce network latency. Utter keeps the entire pipeline completely local and private on your Mac:

• Local Inference: Uses whisper.cpp via native Metal acceleration (ggml-small.en-q5_1).
• Real-Time Audio: 16kHz PCM audio stream via AVAudioEngine, with CoreAudio system taps for transcribing virtual meetings (Zoom, Teams, FaceTime) with speaker separation.
• Fast Mode: Includes an optional instant local engine (Whistle) for zero-latency typing.
• HUD & Island: A non-intrusive floating listening HUD, plus an island buffer so words are not lost if no text box is focused.
• 100% Offline & Private: Zero accounts, zero tracking, zero cloud roundtrips.

Thank you to Georgi Gerganov and the whisper.cpp project for the incredible library that makes fast on-device inference possible on Apple Silicon!

• GitHub Repository: https://github.com/h3x-levi/Utter
• Latest v1.0.0 Release: https://github.com/h3x-levi/Utter/releases/tag/v1.0.0